Algorithm recognition based on demand-driven data-flow analysis

Algorithm recognition is an important problem in program analysis, optimization, and more particularly program comprehension. Basically, one would like to submit a piece of code, and get an answer like "Lines 10 to 23 are an implementation of Gauss-Jordan elimination". Existing approaches often perform a bottom-up recognition, using a database describing many implementation variations of a given algorithm, in a format making them difficult to maintain. In this paper we present a new method to recognize algorithm templates in a program. We perform a top-down recognition, stopping at template variables and thus limiting the need for many program variations. The code to analyze is given in SSA form, and the phi-junctions are computed on-demand, when computation is possible, along the candidate slices.